Predictive Vehicle Diagnostics Using Deep Learning
AI-powered proactive engine fault detection using BiLSTM with Attention on OBD-II telemetry, deployed as a production REST API achieving 90.5% fault recall.
Problem Statement
Modern vehicles generate continuous OBD-II telemetry (RPM, temperature, pressure readings) but consumer systems only provide reactive fault codes after failure occurs. There is no accessible, low-cost predictive system capable of detecting abnormal engine behavior before critical failure using standard OBD-II protocols.
π‘ Key Innovations
- β’ Unsupervised Learning: Autoencoder eliminates need for expensive labeled fault data
- β’ Temporal Modeling: BiLSTM with attention captures patterns across 120-second windows
- β’ 3-Tier Diagnostics: Intelligent fault categorization with actionable recommendations
- β’ Low-Cost Hardware: Compatible with ELM327 OBD-II adapters (<$20)
- β’ Production Ready: 99.9% uptime, <2s API response time
π Dataset
- β’ 1,699 driving sequences (120 seconds each)
- β’ 10 OBD-II sensors sampled at 10Hz (1,200 timesteps)
- β’ Sensors monitored: Coolant Temp, MAP, RPM, Speed, Intake Temp, MAF, Throttle, Ambient Temp, APP_D, APP_E
- β’ Data source: Real-world Seat Leon 1.2 TSI vehicle telemetry
- β’ Split: 1,334 normal / 365 fault sequences (trip-based)
π§ System Architecture
1,699 sequences β’ 1,200 timesteps β’ MinMax normalization β’ Trip-based split
Unsupervised anomaly detection β’ Validation loss: 0.01095
1,334 normal / 365 fault β’ Zero manual labeling
Encoder + Attention + Dense β’ Binary fault prediction
Grid search (62 thresholds) β’ F2-optimized: 0.52
Docker + Flask + Gunicorn β Render.com
πΊοΈ Full Pipeline Workflow Diagram
End-to-end pipeline from raw OBD-II telemetry collection through autoencoder-based anomaly labeling, BiLSTM classification, threshold optimization (F2-maximized), and production deployment on Render.com.
Figure 1: Complete 7-Phase Pipeline β Data Collection β Preprocessing β Autoencoder β Auto-Labeling β BiLSTM Classifier β Threshold Optimization β Production Deployment
π Model Performance
β’ Diagnostic Capabilities: 3-tier system (Precise β General β Safety)
β’ Fault Categories: 8 categories (Sensor, Thermal, Ignition, Throttle, RPM, Electrical)
β’ Sensor Identification: Attention mechanism highlights critical sensors
β’ Cost Savings: $225 per test cycle (false alarm reduction)
π Live Deployment
API Base URL:
https://pbl-predictivevehiclediagnostics.onrender.com
β’ GET /health β System status & threshold
β’ POST /predict β Fault probability + diagnostic tier + sensors
β’ POST /batch_predict β Process multiple sequences
π οΈ Technology Stack
Machine Learning
Backend
Deployment
Model Architecture: BiLSTM with Attention Mechanism | Input: (1200, 10) | Output: Binary fault probability
π Repository
DemWrath/PBL_PredictiveVehicleDiagnosticsclassifier.h5 - Trained model (2.75 MB)
deploy_cloud_api.py - Flask API server
diagnostic_engine.py - 3-tier diagnostics
lstm_classifier.py - Model architecture
config.py - Hyperparameters
Dockerfile - Container spec
π Future Enhancements
Phase 4.5: Enhanced Diagnostics
- β’ 37 specific fault pattern detection (sensor-level)
- β’ Physics-based validation rules
- β’ Repair cost estimation engine
Phase 7: Real-World Testing
- β’ Deploy to test fleet (5-10 vehicles)
- β’ Collect production fault data
- β’ Retrain with real failure examples
Phase 8: Mobile Application
- β’ iOS/Android companion app
- β’ Real-time push notifications
- β’ OBD-II adapter pairing (Bluetooth)
- β’ Maintenance schedule integration
π Academic Credits
Project Guide
Dr. Neha Janu
Student
Abhinav Jha
ID: 2427030088
Institution
Manipal University Jaipur
Dept. of CSE β’ 2026
Β© 2026 PBL Project | Predictive Vehicle Diagnostics Using Deep Learning
View Source Code |
API Status